R 中的多结果逻辑回归

Question

This question was migrated from Stack Overflow because it can be answered on Cross Validated. Migrated 5 days ago .

I have a large dataset with multiple binary dependent variables (eg y1, y2, y3, y4). yi variables are people's response to their decision making behaviour in different situations. There are many explanatory variables (eg x1, x2, ...) which are mainly sociodemographic variables.

I can model each y based on the explanatory variables (eg y1~x).

But, all binary dependent variables are correlated and I would like their correlations to be considered in my model. In other words, I would like to model yi ~ xj for all i simultaneously. It can be called a multiple outcome logistic regression model.

I have already tried looking up on the.net, glm, and brms packages. None worked for me.

Does anybody know a package that is capable of performing such model?

Answer 1

I think I've figured it out:

Make sure that the installer project includes the Content Files for the web app. Then go into the web app's csproj file with a text editor and find the ItemGroup where the static Content files are listed. Then add an entry with a wildcard in it and the installer will package them in.

<ItemGroup>
    <Content Include="Content\test.html" />
    <Content Include="Content\*.html" />
        <Content Include="Content\*.gif" />
       <Content Include="Content\*.css" />
       <Content Include="Content\*.jpg" />
       <Content Include="Content\*\*.js" />

Answer 2

Right click on the folder in the File System view of the Web Setup project, and select Add, File...

It doesn't look like you can add an entire folder this way though, so you would have to add every file individually to the Setup program.

EDIT:

Aren't all files from your project folder included in a WebSite project in Visual Studio? Then it's just a case of having your WebSite project and Web Setup project in the same solution, then right-click on the Web Application Folder in the Web Setup project, select Add, Project Output... and then select the WebSite Project in the drop down and Content Files from the list box and click OK.

Answer 3

Maybe .net::multinom for multinominal logistic regression fits your problem.

If your dichotomous dependent variables are not from a categorical one, perhaps a structural equation model with correlated dichotomous dependent variables is a solution. That would work with lavaan .

Answer 4

There are two techniques I would like to suggest. When you have a matrix $Y$ of outcomes and one of predictors $X$ , you can use the Canonical Correlation Analysis, implemented in the CCA package in R. This technique searches for new columns that represent the best linear combinations between $Y$ and $X$ . Imagine, you can compute the correlation between two variables and that tells you something about their relationship, but you can't compute the correlation between two sets of columns*! The CCA finds new columns $a_{Y}$ and $b_{X}$ that express the highest correlation between $X$ and $Y$ , accounting for the variability within and between the two matrices. Here there is not really a relationship outcomes-predictors, because $X$ and $Y$ are considered at the "same level". You can also use the Partial Least Square (PLS) analysis, implemented in the pls package. This technique is intimately linked to the CCA, but in this case formally you are looking for the best approximations of $X$ based on their correlation with $Y$ . I won't get in further details for now, but if you need more clarifications please ask.

Here is a tutorial for the CCA. Here another one.

Here is a tutorial for the PLS. Here another one.

*You can compute the correlation of one variable with each of the other ones (pair-wise correlations), but this doesn't take in account the correlation with the other variables!

Answer 5

Unfortunately I don't think there is an out of the box way to do this, as projects maintains references to the files individually.

You could achieve what you need via a custom macro, or asking your other team to edit the .csproj XML file (perhaps simpler or riskier, depending on their background).